Workflow run

Final post

SAP published an architectural breakdown of Joule. It's the cleanest public map yet of how Joule actually works under the hood — runtime, skill execution, grounding mechanisms.

Worth reading. And worth reading for what it doesn't address.

The post does two things well: it explains how Joule answers questions in natural language, and how it grounds those answers against SAP data through skills wired into S/4HANA, SuccessFactors, and Ariba.

What it doesn't address — because Joule wasn't designed for it — is the third layer. Execution.

Three layers of AI in SAP, and they are not the same job:

1. Answer layer. Natural language interaction. Ask about an open delivery in VL06O, get a readable summary. Joule does this well.

2. Retrieval layer. Grounding the answer against SAP data, skills, embeddings, all under SAP governance. Joule does this too.

3. Execution layer. Build a missing report column, adapt a workflow, trigger an ME22N update, change a delivery block, push the operational decision back into SAP itself.

The execution layer is where the conversation usually goes sideways. Once the business sees Joule answer a question, the next ask is: "great, now do something about it." That's a different architecture problem.

You can't bolt execution onto an answer layer. It requires operational context, authorizations enforced at the field and record level, transport-safe deployment, and the ability to build inside SAP without a full ABAP cycle for every change.

That's an operational layer, not an assistant. Different category, different job.

Joule is a real piece of SAP's AI story. It's also not the whole story — and the architects reading that post will see the gap before anyone else does.

Tagging Timo Elliott and Joshua Greenbaum — would value your read on where the execution-layer conversation is heading inside SAP shops under S/4 pressure.

Comment "GAP" and I'll send the architecture map we're walking SAP teams through.